Sparse spectrally arbitrary patterns

Electronic Journal of Linear Algebra Volume 28 Volume 28: Special volume for Proceedings of Graph Theory, Matrix Theory and Interactions Conference Article 8 2015 Sparse spectrally arbitrary patterns Brydon Eastman Redeemer University College, beastman@redeemerca Bryan L Shader University of Wyoming, bshader@uwyoedu Kevin N Vander Meulen Redeemer University College, kvanderm@redeemerca Follow this and additional works at: http://repositoryuwyoedu/ela Recommended Citation Eastman, Brydon; Shader, Bryan L; and Vander Meulen, Kevin N (2015), "Sparse spectrally arbitrary patterns", Electronic Journal of Linear Algebra, Volume 28 DOI: https://doiorg/1013001/1081-38103010 This Article is brought to you for free and open access by Wyoming Scholars Repository It has been accepted for inclusion in Electronic Journal of Linear Algebra by an authorized editor of Wyoming Scholars Repository For more information, please contact scholcom@uwyoedu

SPARSE SPECTRALLY ARBITRARY PATTERNS BRYDON EASTMAN, BRYAN SHADER, AND KEVIN N VANDER MEULEN Abstract We explore combinatorial matrix patterns of order n for which some matrix entries are necessarily nonzero, some entries are zero, and some are arbitrary In particular, we are interested in when the pattern allows any monic characteristic polynomial with real coefficients, that is, when the pattern is spectrally arbitrary We describe some order n patterns that are spectrally arbitrary We show that each superpattern of a sparse companion matrix pattern is spectrally arbitrary We determine all the minimal spectrally arbitrary patterns of order 2 and 3 Finally, we demonstrate that there exist spectrally arbitrary patterns for which the nilpotent-jacobian method fails Key words Companion matrix pattern, Spectrally arbitrary pattern, Nilpotent-Jacobian method, Nonzero pattern AMS subject classifications 15A18, 15B99, 05C50 1 Introduction In this paper, a pattern of order n is an n-by-n matrix with entries in {0,, } The qualitative class of a pattern A = [A i,j ], denoted by Q(A), is the set of all real matrices A = [A i,j ] such that A i,j 0 if A i,j =, A i,j = 0 if A i,j = 0, and A i,j is unrestricted if A i,j = We say A,ij is nonzero if A i,j {, } While the notation # has been used in the literature (see eg [1, 6]) instead of, we find the latter more visibly descriptive A pattern A realizes a polynomial p(x) if there is a matrix A Q(A) such that the characteristic polynomial of A is p(x) A pattern A is spectrally arbitrary if A realizes every monic polynomial of degree n with real coefficients For example the Received by the editors on September 22, 2014 Accepted for publication on April 15, 2015 Handling Editor: Stephen J Kirkland Department of Mathematics, Redeemer University College, Ancaster, ON, Canada, L9K 1J4 (beastman@redeemerca, kvanderm@redeemerca) Research supported in part by NSERC Department of Mathematics, University of Wyoming, Laramie, WY 82071-3036 (bshader@uwyoedu) 83

84 B Eastman, B Shader, and KN Vander Meulen order n pattern 0 0 0 C n = 0 0 0 is spectrally arbitrary since it is the pattern of a companion matrix The concept of spectrally arbitrary pattern was introduced in [4] in the context of sign patterns (whose entries are in {0, +, }) and have since been explored by various other authors Spectrally arbitrary zero-nonzero patterns (whose entries are restricted to {0, }) have also been explored; see for example [2] More recently, Cavers and Fallat [1] have developed spectral results for more general classes of combinatorial patterns A pattern is reducible if it is permutationally equivalent to a block triangular matrix pattern and irreducible otherwise Note that if A is irreducible, it is possible that there is a reducible matrix A Q(A) Cavers and Fallat [1] noted that an irreducible spectrally arbitrary pattern of order n has at least 2n 1 nonzero entries Throughout this paper, we call an order n spectrally arbitrary pattern sparse if it has exactly 2n 1 nonzero entries In Section 3, we characterize all the order 2 and 3 patterns that are spectrally arbitrary In Section 3, we also introduce a spectrally arbitrary pattern for which a commonly-used method, the nilpotent-jacobian method, fails In Section 4, we answer a question raised by [1, Question 44] In particular, we show that there exist sparse irreducible spectrally arbitrary patterns that have at least two nonzero entries on the main diagonal A pattern B is a superpattern of pattern A if B can be obtained from A by replacing some (or none) of the zero entries of A with A pattern B is a relaxation of A if B can be obtained from A by changing some of the entries of A to and/or some of the zero entries of A to Note that in [1], the definition of superpattern is slightly different; it includes some relaxations of the superpatterns In particular, an entry of the superpattern B could be if A is 0 in [1] If A is spectrally arbitrary, then every relaxation of A is also spectrally arbitrary (see [1, Lemma 21]) It was shown in [1] that every superpattern of the companion pattern C n is spectrally arbitrary Recently [5], the class of all sparse companion matrix patterns were classified, and it was noted that some superpatterns of these patterns are spectrally arbitrary In Section 2, we demonstrate that all superpatterns of the sparse companion patterns are spectrally arbitrary

Sparse spectrally arbitrary patterns 85 2 Techniques for spectrally arbitrary patterns A common technique for showing a pattern is spectrally arbitrary, introduced in [4] for sign patterns, is the nilpotent-jacobian method This method was extended in [1, Theorem 211] to apply to other classes of patterns such as those discussed in this paper The nilpotent-jacobian method Let N be a pattern of order n with m n nonzero entries and N be a nilpotent matrix in Q(N ) Among the m nonzero entries, choose n nonzero entries, say N i1,j 1,, N in,j n and let A be the matrix obtained from N by replacing the entries N i1,j 1,, N in,j n by real variables a 1,, a n If the characteristic polynomial of A is p A (x) = x n + p 1 x n 1 + p 2 x n 2 + + p n 1 x + p n, let Jac(N) be the Jacobian matrix of order n with (i, j) entry equal to pi a j evaluated at (a 1,, a n ) = (N i1,j 1,, N in,j n ) If Jac(N) has rank n, then every superpattern of N, including N itself, is spectrally arbitrary To simplify the wording, we say that N allows a nonzero Jacobian if Jac(N) has rank n for some nilpotent matrix N Q(N ) and choice of variables Thus the nilpotent-jacobian method can be summarized as: if N allows a nonzero Jacobian, then every superpattern of N is spectrally arbitrary It is not known whether every zero-nonzero spectrally arbitrary pattern allows a nonzero Jacobian over R, but it was shown in [8] that there exist spectrally arbitrary zero-nonzero patterns that do not allow a nonzero Jacobian over C In Section 3, and again in Section 4, we demonstrate that not every spectrally arbitrary {0,, }-pattern allows a nonzero Jacobian over R The nilpotent-jacobian method is an analytic method for determining if a pattern is spectrally arbitrary In [7], an algebraic method, called the nilpotent-centralizer method, was introduced for sign patterns Taking appropriate account of the positions, this method can also be extended to {0,, }-patterns as described below We use A B to represent the Hadamard, or the entry-wise product, of matrices A and B of the same size If A is a matrix and B is a pattern, then an entry of A B is nonzero if and only if the corresponding entries in both A and B are nonzero Note that the index of a nilpotent matrix N is the smallest positive integer k such that N k = O Based on Theorem 35 and Lemma 36 in [7], we have the following lemma which implies the validity of the subsequent nilpotent-centralizer method Lemma 21 Let N Q(N ) be a nilpotent matrix of order n Then Jac (N) has rank n if and only if N has index n and the only matrix B in the centralizer of N satisfying B T N = O is the zero matrix The nilpotent-centralizer method Let N Q(N ) be an order n nilpotent matrix of index n If the only matrix B in the centralizer of N satisfying B T N = O is the zero matrix, then the pattern N and each of its superpatterns is spectrally arbitrary pattern

86 B Eastman, B Shader, and KN Vander Meulen Two patterns A and B are equivalent if B can be obtained from A by transposition and/or permutation similarity In [1, Example 212], using a pattern equivalent to C n, the nilpotent-jacobian method is used to demonstrate that every superpattern of C n is spectrally arbitrary Below we use the nilpotent-centralizer method to get the same result for a larger class of matrices that includes C n For 1 k n 1, we say the k th subdiagonal of a pattern is the set of positions {(i, i k) : k + 1 i n} A family of patterns C n was introduced in [5]: a pattern A is in C n if A has an entry in each superdiagonal position, exactly one along the main diagonal, exactly one on each subdiagonal, and zeros elsewhere For example, the pattern C n is in C n In [5], it was observed that the patterns in C n characterize those patterns that uniquely realize each characteristic polynomial up to diagonal similarity We now show that each superpattern of a pattern in C n is spectrally arbitrary Theorem 22 Let A be a pattern in C n Then every superpattern of A is spectrally arbitrary Proof Let N be in C n Let N Q(N ) be the matrix with ones on the superdiagonal and zeros elsewhere Note that N is nilpotent of index n Let B be a matrix in the centralizer of N such that B T N = O Note that since B is in the centralizer of N, B = q(n) for some polynomial q(x) of degree at most n 1 Let k {1,, n 1} Observe that the transpose of N n k has exactly k nonzero entries: it has k ones on the (n k) th subdiagonal, and zeros elsewhere But N has a entry on each subdiagonal, and hence the coefficient of N n k must be zero in q(n) since B T N = O It follows that q(x) is a constant But N has a nonzero entry on the main diagonal, hence q(x) must be the zero polynomial Thus every superpattern of N is spectrally arbitrary by the nilpotent-centralizer method Suppose that A and B are order n patterns, each with only one nonzero element in the main diagonal, A k,k 0 and B l,l 0 for some l k Then B is said to be a diagonal slide of A if B i,j = A i,j for all i j In the next argument, we use the fact that when N is an order n nilpotent matrix of index n, then the matrix B is in the centralizer of N if and only if B = q(n) for some polynomial q(x) Theorem 23 Suppose A is a spectrally arbitrary pattern with only one nonzero entry on the main diagonal and A allows a nonzero Jacobian If B is a diagonal slide of A, then any superpattern of B is spectrally arbitrary Proof Let A be an order n pattern, and suppose that B is a diagonal slide of A and N Q(A) is a nilpotent matrix such that Jac(N) has rank n Then by Lemma 21, N has index n and if p(x) is a polynomial of degree at most n such that p(n) A T = O, then p(x) is the zero polynomial

Sparse spectrally arbitrary patterns 87 Since N is nilpotent and A has only one nonzero entry on the main diagonal, N has all zeros on the main diagonal Since B is a diagonal slide of A and N has all zeros on its diagonal, N is a nilpotent realization of B Suppose q(x) is a polynomial of degree at most n such that q(n) B T = O Then for some constant c, we have (q(n) ci) A T = O, and thus p(x) = q(x) c is the zero polynomial But then q(x) = c, and q(n) B T = O implies that c = 0 Thus q(x) = 0 and, by the nilpotent-centralizer method, any superpattern of B is spectrally arbitrary The pattern C n is an example of a pattern that preserves the property of being spectrally arbitrary under a diagonal slide, since each pattern in C n is spectrally arbitrary as illustrated in Theorem 22 In fact, Theorem 22 implies that C n preserves the property under any subdiagonal slide as well (where a subdiagonal slide involves moving a lone entry on the k th subdiagonal to another position on the same subdiagonal, for some k {1,, n 1}) In Section 4, we have an example of a spectrally arbitrary pattern, that does not allow a nonzero Jacobian, for which a subdiagonal slide restricts the possible characteristic polynomials In particular, moving entry (4, 1) of the spectrally arbitrary pattern Y 6 (6, 2) to position (5, 2) produces a pattern that requires singularity (see Section 4 for the definition of Y n (n, k)) We do not know if the conclusion of Theorem 23 will still hold if we drop the condition of allowing a nonzero Jacobian We do have examples of spectrally arbitrary patterns that do not allow a nonzero Jacobian that preserve the property of being spectrally arbitrary under a diagonal slide: F in Section 3 (see Theorem 38) and Y n (s, k) in Section 4 (see Theorem 44 and 47) 3 Spectrally arbitrary patterns of order 2 and 3 In this section we characterize all the patterns of order 2 and 3 that are spectrally arbitrary up to equivalence A tool that is used to classify permutationally equivalent patterns is the digraph of the pattern Given an order n matrix pattern A, the digraph D(A) has vertex set {v 1, v 2,, v n } with arc set {(v i, v j ) : A i,j 0} Two patterns are permutationally equivalent if and only if their labeled digraphs are isomorphic (labeling arcs with and as appropriate) Further, the digraph of the transpose of A is obtained from D(A) by reversing all the arcs For 1 k n, a cycle of length k, or k-cycle, in a digraph D is a sequence of k arcs (v i1, v i2 ), (v i2, v i3 ),, (v ik, v i1 ) with k distinct vertices v i1, v i2,, v ik A 1-cycle is often called a loop of D Lemma 31 [1, Corollary 28] If A is a spectrally arbitrary pattern, then the digraph D(A) has at least one loop and at least one 2-cycle If D(A) has cycle (v i1, v i2 ), (v i2, v i3 ),, (v ik, v i1 ) and A = [A i,j ] A then the associated cycle product is ( 1) k 1 A vi1,v i2 A vi2,v i3 A vik,v i1 A composite cycle of length k is a set of vertex disjoint cycles with lengths summing to k (with an associated cycle product being the product of the cycle products of the individual cycles) If a

88 B Eastman, B Shader, and KN Vander Meulen k-cycle is not composite, then we say the k-cycle is a proper k-cycle We will use the following fact (see for example [6]): Lemma 32 Given A Q(A), if f k is the sum of all the composite cycle products of length k in D(A), then the characteristic polynomial of A is p A (x) = x n f 1 x n 1 + f 2 x n 2 + + ( 1) n f n By Lemma 32, if the digraph of a pattern A has exactly one k-cycle for some k, 1 k n, and A has only entries in the positions corresponding to the arcs of the k-cycle, then that pattern does not realize a polynomial with f k = 0 Thus we have the following lemma: Lemma 33 Suppose k {1, 2,, n} If A is a spectrally arbitrary pattern of order n, and D(A) has exactly one composite cycle of length k, then that cycle must have an arc labeled The next result is a {0,, }-pattern version of a corresponding sign pattern result [1, Theorem 32] Theorem 34 If A is a spectrally arbitrary pattern of order 2 then A is, up to equivalence, a relaxation of either C 2 = [ 0 ] [ or T 2 = ] The following result was stated for sign patterns in [3, Proposition 21], but the proof did not rely on the signs Theorem 35 The direct sum of patterns of which at least two are of odd order is not spectrally arbitrary Furthermore if the direct sum of spectrally arbitrary patterns has at most one odd order summand, then the direct sum is spectrally arbitrary A pattern B is a minimal spectrally arbitrary pattern if A is not spectrally arbitrary for any A B such that B is a superpattern of A or B is a relaxation of A The following theorem follows from Theorems 34 and 35 and demonstrates that, up to equivalence, the reducible, minimal spectrally arbitrary patterns of order 3 are T 2 [ ] and C 2 [ ] Theorem 36 If R is a reducible spectrally arbitrary pattern of order 3 then R is equivalent to either a superpattern or a relaxation of T 2 [ ] or C 2 [ ] Next, we next explore irreducible patterns

Sparse spectrally arbitrary patterns 89 Lemma 37 [2, Theorem 11] If A is an irreducible, minimal spectrally arbitrary zero-nonzero pattern of order 3, then A is equivalent to 0 V 3 =, or T 3 = 0 Further, every superpattern of V 3 and T 3 is spectrally arbitrary We will see in Theorem 39 that the following {0,, }-patterns are irreducible, minimal spectrally arbitrary patterns: 0 0 0 C 3 = 0, D =, E =, 0 0 0 0 0 G = 0 0, H = and F = 0 0 0 0, K = We first observe that F is a spectrally arbitrary pattern for which the nilpotent- Jacobian method fails Theorem 38 If A is equivalent to F or a diagonal slide of F, then A is spectrally arbitrary, but A does not allow a nonzero Jacobian Proof Suppose A is a diagonal slide of F and N Q(A) is nilpotent It follows from Lemma 32 that the diagonal position of A will be zero in N Also, since det(n) = 0, at least one of N 1,2 and N 2,3 is zero Considering the 2-cycles of D(A), it follows from Lemma 32 that N 1,2 = N 2,3 = 0 Thus N has zero entries in each position of A Let B be a matrix of order 3 having only zero entries, except B 3,1 = 1 Then B T A = O and B is in the centralizer of N for any nilpotent matrix in Q(A) Thus, by Lemma 21, A does not allow a nonzero Jacobian Suppose r 1, r 2, r 3 R, and p(x) = x 3 + r 1 x 2 + r 2 x + r 3 Let 0 a 0 F = 1 b c Q(F) d 1 0 If r 3 0, let t R\{0, r 2 } and (a, b, c, d) = ( r 2 t, r 1, t,, ) r 3 (r 2+t)t If r 3 = 0, let (a, b, c, d) = ( r 2, r 1, 0, 1) In each case, d is nonzero as required for F, and the characteristic polynomial of F is p(x) Thus F is spectrally arbitrary

90 B Eastman, B Shader, and KN Vander Meulen Next, we will show that any diagonal slide of F is also a spectrally arbitrary Let F be a diagonal slide of F Then either F 1,1 = or F 3,3 = By permutation equivalence, we can assume F 1,1 = Consider the following matrix F = b a 0 1 0 c d 1 0 Q(F ) Suppose { r 1, r 2, r} 3 R, and p(x) ( = x 3 + r 1 x 2 + r 2 x + ) r 3 If r 1 0, let t R\ 0, r 2, r3 r 1 and (a, b, c, d) = r 2 t, r 1, t, tr1+r3 (r 2+t)t If r 1 = 0 and r 3 0, ( ) let t R\ {0, r 2 } and (a, b, c, d) = r 2 t, 0, t, If r 1 = 0 and r 3 = 0, let r 3 (r 2+t)t (a, b, c, d) = (0, 0, r 2, 1) In each case d is nonzero and the characteristic polynomial of F is p(x) Therefore every diagonal slide of F is spectrally arbitrary The next theorem determines the minimal spectrally arbitrary {0,, }-patterns of order 3 Theorem 39 If A is an irreducible pattern of order 3, then A is spectrally arbitrary if and only if, up to equivalence, A is a relaxation of a superpattern of one of the patterns E, G, H, K, V 3 or T 3, or of a diagonal slide of C 3, D, or F Proof By Lemma 37, both V 3 and T 3 are spectrally arbitrary By Theorem 22, every superpattern of a diagonal slide of C 3 is spectrally arbitrary and it follows from Lemma 33 that C 3 is not the relaxation of another spectrally arbitrary pattern Hence any diagonal slide of C 3 is a minimal spectrally arbitrary pattern Consider the matrices a 1 1 0 a 1 1 0 1 1 0 1 1 0 a 1 1 1 a 2 0 1, a 2 a 3 1, a 1 a 2 1, a 1 a 2 1 and a 2 0 1 a 3 1 0 0 1 0 0 a 3 0 a 3 0 0 a 3 2 0 Setting the variables (a 1, a 2, a 3 ) as (0, 1, 0), (0, 1, 0), ( 1, 1, 0), ( 1, 1, 0), and (0, 2, 4) respectively, for each pattern D, E, G, H, and K, we obtain a nilpotent matrix for which the corresponding Jacobian has full rank Thus the superpatterns of D, E, G, H, and K are spectrally arbitrary by the nilpotent-jacobian method and, by Theorem 23, any superpattern of a diagonal slide of D is spectrally arbitrary By Theorem 38 we know every diagonal slide of F is spectrally arbitrary We next show that all of their superpatterns are also spectrally arbitrary Let B be a superpattern of F If B 1,3 0 then B is equivalent to a relaxation of K via the permutation (12) If B 1,1 0 then B T is a relaxation of a superpattern of G If B 3,3 0 then B is equivalent, via transpose and the permutation (13), to a relaxation of a superpattern of V 3 Thus any superpattern of F is spectrally arbitrary

Sparse spectrally arbitrary patterns 91 Now let B be a superpattern of a diagonal slide of F By transpose and the permutation (13), we can assume B 1,1 = If B 1,3 0 then B is a relaxation of K If B 2,2 0 then B T is a relaxation of a superpattern of G If B 3,3 0 then B is a relaxation of a superpattern of V 3 Thus any superpattern of B is also spectrally arbitrary Therefore any superpattern of a diagonal slide of F is spectrally arbitrary Suppose A = [a i,j ] is an irreducible, minimal spectrally arbitrary pattern of order 3 As noted in the introduction, a spectrally arbitrary pattern of order n must have at least 2n 1 nonzero entries Thus A must have at least five nonzero entries If A has no entries, then by Lemma 37, A is equivalent to V 3 or T 3 So suppose A has at least one entry By Lemma 31 and the fact that A is irreducible, A must be equivalent to a relaxation of a superpattern of V 1 = 0 0 0 0 We consider the following cases:, or V 2 = 0 0 0 0 1 Suppose A has exactly five nonzero entries By Lemmas 31 and 33, one of the nonzero entries must be a on the main diagonal If A is a relaxation of V 2 and not V 1 then A does not realize the polynomial x 3 +1, regardless of the position of the diagonal entry Thus A is a relaxation of V 1 By Lemma 33, either a 1,2 = or a 2,1 = If a 1,2 = and a 2,1 =, then A does not realize the polynomial x 3 + 1 Thus a 2,1 = If a 2,3 = a 3,1 =, then A does not realize the polynomial x 3 + x Thus a 3,1 = or a 2,3 = If a 3,1 =, then A is a diagonal slide of a relaxation of C 3 If a 2,3 =, then A T is equivalent to a diagonal slide of a relaxation of C 3, via permutation (12) 2 Suppose A has exactly six nonzero entries and A is a relaxation of a superpattern of V 1 with only one entry on the main diagonal Thus exactly one of a 3,2 and a 1,3 is not zero But the case with a 1,3 0 is equivalent to the former via permutation (123) Thus assume a 3,2 0 In order to realize the polynomial x 3, either a 3,1 =, a 1,2 =, or a 2,3 = (a) Suppose a 3,1 = Then A is a relaxation of a diagonal slide of D (b) Suppose a 1,2 = and a 3,1 = If a 2,3 = a 3,2 = then A does realize characteristic polynomial x 3 Thus either a 2,3 = or a 3,2 = If a 3,2 =, then A would be equivalent to a superpattern of a diagonal slide of C 3 via permutation (132) Thus a 2,3 =, and A is a diagonal slide of F (c) Suppose a 2,3 = and a 3,1 = a 1,2 = In order for A to realize a characteristic polynomial x 3, it is necessary that a 2,1 = Therefore A is equivalent to a superpattern of a diagonal slide of C 3 via transposition and permutation (12)

92 B Eastman, B Shader, and KN Vander Meulen 3 Suppose A has exactly six nonzero entries with two nonzero entries on the main diagonal and A is a relaxation of a superpattern of V 1 If a 1,1 0 and a 3,3 0 then A is a relaxation of V 3 If a 2,2 0 and a 3,3 0 then A is equivalent to a relaxation of V 3 via transposition and permutation (12) Suppose that a 1,1 0 and a 2,2 0 By Lemma 33 either a 3,1 =, a 1,2 =, or a 2,3 = If a 1,2 = and a 3,1 = a 2,3 =, then A does not realize characteristic polynomial x 3 + x Thus a 3,1 = or a 2,3 = If a 3,1 =, then A is a relaxation of H If a 2,3 =, then A is equivalent to a relaxation of H via transpose and permutation (12) 4 Suppose A is a relaxation of a superpattern of V 2 but not of V 1, and A has six nonzero entries Then A has two nonzero entries on the main diagonal If a 1,1 0 and a 3,3 0, then A is a relaxation of T 3 The case with a 1,1 0 and a 2,2 0 is equivalent, via permutation (13), to the case with a 2,2 0 and a 3,3 0 Suppose a 1,1 and a 2,2 are nonzero By Lemma 33, a 3,2 =, a 2,3 = or a 1,1 = (a) Suppose a 3,2 = Then A is a relaxation of G (b) Suppose a 2,3 = Then A T is a relaxation of G (c) Suppose a 1,1 = and a 3,2 = a 2,3 = Then a 2,2 =, otherwise A does not realize characteristic polynomial x 3 Thus A is a relaxation of E 5 Suppose A has seven nonzero entries Since A is irreducible, up to equivalence, A is a relaxation of one of the five patterns:,,, 0 and The first and second patterns are superpatterns of V 3 The third and fourth patterns are superpatterns of T 3 Suppose that A is a relaxation of the fifth pattern By Lemma 33 the diagonal entry must be a and thus A is a relaxation of K Note that any pattern with more than 7 nonzero entries is equivalent to a relaxation of a superpattern of V 3 Corollary 310 Each named pattern in Theorem 39 is a minimal spectrally arbitrary pattern We conclude this section by noting a consequence for {0, }-patterns that follows from the fact that every {0, }-pattern is a {0,, }-pattern Corollary 311 If A is an irreducible spectrally arbitrary {0, }-pattern of

Sparse spectrally arbitrary patterns 93 order 3, then A is equivalent to a relaxation of a diagonal slide of 0 0, 0 0 or a relaxation of one of 0 0 0 or 0 0 0 4 A class of spectrally arbitrary patterns that do not allow a nonzero Jacobian It is unknown (see for example [9]) whether an irreducible zero-nonzero spectrally arbitrary pattern must allow a nonzero Jacobian As we saw in the previous section, the pattern F is an example of an irreducible spectrally arbitrary {0,, }- pattern that does not allow a nonzero Jacobian In this section we demonstrate that F is not a solitary exception; we introduce a whole class of irreducible spectrally arbitrary {0,, }-patterns that do not allow a nonzero Jacobian We also answer Question 44 raised in [1] in the affirmative with Corollary 46 below Given integers k, s and n with 1 k < s n, let Y n (s, k) be the lower Hessenberg pattern of order n with a in each superdiagonal position, a in position (i, 1) for all i s, 1 i n and a in position (s, s k + 1) Note there are two entries on the (k 1) st subdiagonal in this pattern We will see that this pattern does not allow a nonzero Jacobian Nevertheless, it can be shown, by examining characteristic polynomials, that for certain choices of s, k, and n, the pattern is spectrally arbitrary Any matrix in Q(Y n (s, k)) is diagonally similar to an order n unit Hessenberg matrix y 1 1 0 0 0 y k 0 Y n(s, k) = y s 1 0 0 t y s+1 0 0 1 y n 0 0 0

94 B Eastman, B Shader, and KN Vander Meulen with y i, t R Note that the digraph of Y n (s, k) has two k-cycles, no s-cycle, and exactly one l-cycle for each l {s, k}, 1 l n The only composite l-cycles that are not proper l-cycles involve the entry t in position (s, s k + 1) The following description of the characteristic polynomial of Y n (s, k) then follows from Lemma 32 Lemma 41 Let y s = 0 and y 0 = 1 The characteristic polynomial of Y = Y n (s, k) is p Y = x n n y i x n i + i=1 s (ty i k )x n i i=k Lemma 42 arbitrary If s is not a multiple of k, then Y = Y n (s, k) is not spectrally Proof Suppose s is not a multiple of k, and the characteristic polynomial of Y Q(Y) is x n + x n s By Lemma 41, y i = 0 for 1 i < k Note that there exists an integer c such that 1 s ck < k Since y s ck = 0, we can show by induction on h that y s (c h)k = 0 for h {1, 2,, c 1} In particular, y s k = 0 But the coefficient of x n s in p Y is y s k t Thus Y cannot realize the characteristic polynomial x n + x n s Therefore Y is not spectrally arbitrary Lemma 43 Suppose s is a multiple of k If Y is a nilpotent realization of Y = Y n (s, k), then Y ij = 0 whenever Y ij = Proof Suppose Y = Y n (s, k) is nilpotent, and t 0 Let p Y be the characteristic polynomial of Y as in Lemma 41 Since Y is nilpotent, y i = 0 for all i {k,, s}, 1 i n Thus the coefficient of x n s in p Y is ty s k Hence y s k = 0 We can show, by induction on c, that y s ck = 0 for c {1,, s k k } Therefore the coefficient of x n k is t 0 But this contradicts the fact that Y is nilpotent Therefore t = 0 By Lemma 41, it follows that y i = 0 for all i since Y is nilpotent Theorem 44 Suppose s is a multiple of k The pattern Y n (s, k) does not allow a nonzero Jacobian Proof Suppose Y = Y n (s, k) is nilpotent By Lemma 43, Y is a unit Hessenberg matrix with all entries zero except those on the superdiagonal Thus Y has index n Let B be the matrix such that B T has ones on the (s 1) th subdiagonal and zeros elsewhere Then B T Y = O and BY = Y B Thus by Lemma 21, Y does not allow a nonzero Jacobian Theorem 45 For any n > 2, the pattern Y n (s, k) is a (minimal) spectrally arbitrary pattern if and only if s is an odd multiple of k Proof Let Y = Y n (s, k) By Lemma 42, we may assume s = ck for some positive

Sparse spectrally arbitrary patterns 95 integer c Let p(x) = x n + r 1 x n 1 + r 2 x n 2 + + r n 1 x + r n for some r 1,, r n R It is enough to show that p Y (x) = p(x) for some y 1,, y s 1, t, y s+1,, y n R if and only if c is odd By Lemma 41, we seek a real solution to the system of equations: r 1 = y 1 r k 1 = y k 1 r k = t y k r k+1 = ty 1 y k+1 (41) r s 1 = ty s k 1 y s 1 r s = ty s k r s+1 = y s+1 r n = y n It follows from the above that y q = r q for all q {1,, k 1, s + 1,, n} Setting r 0 = 1 and using forward substitution, we can solve for y m for each m {k, k + 1,, s 1}, starting at m = k, to obtain m k y m = r m ki t i After making these substitutions, the s th equation in (41) yields r s = i=0 c r s ki t i (42) i=1 Suppose c is even Then the characteristic polynomial of Y can not be x n + x n s since if r 1 = = r s 1 = 0, then (42) would imply r s = t c < 0 Thus, if c is even, Y n (s, k) is not spectrally arbitrary Suppose c is odd Then we let t be a real root of the monic polynomial c i=0 r s kit i of degree c It follows that Y = Y n (s, k) can achieve any characteristic polynomial Since Y is sparse, there is no spectrally arbitrary pattern B such that Y is a proper superpattern of B By Lemma 43, there exists no spectrally arbitrary B such that Y is a relaxation of B Therefore Y n (s, k) is a minimal spectrally arbitrary pattern When k = 1, then pattern Y n (s, k) has two nonzero entries on the main diagonal

96 B Eastman, B Shader, and KN Vander Meulen For example, 0 0 Y 4 (3, 1) = 0 0 0 0 0 0 0 Further, each pattern Y n (s, k) with s < n is irreducible Thus by Theorem 45, for odd s < n, Y n (s, 1) is a sparse spectrally arbitrary pattern that has the characteristics sought in [1, Question 44]: Corollary 46 For each n 4, there exists a sparse irreducible spectrally arbitrary pattern that has two nonzero entries on the main diagonal We finish this section by observing that other spectrally arbitrary patterns can be obtained from Y = Y n (s, k), for example, by using a diagonal or subdiagonal slide In the proof of Theorem 45, for any particular polynomial p(x), we determined entries y i in Y = Y n (s, k) by forward substitutions, while solving for t depended on the polynomial equation in (42), so that p Y (x) = p(x) We will show that a diagonal slide of Y will change the coefficients in the characteristic polynomial of Y in such a way that one can still solve for each y i via forward substitution and the corresponding equation to (42) will still be a monic polynomial in t of odd degree c = s k Theorem 47 Let s be an odd multiple of k > 1 Then any diagonal slide of Y n (s, k) is a minimal spectrally arbitrary pattern Proof Let k > 1 Suppose B is a diagonal slide of Y n (s, k) such that B l,l = for some 1 l n Consider Y = Y n (s, k) and let B Q(B) be a matrix such B l,l = y 1 and B i,j = Y i,j for all i j Note the digraph of B has two proper k- cycles (since s > 2k), no proper s-cycle, and exactly one proper m-cycle for each m {s, k}, 1 m n The composite m-cycles that are not proper m-cycles involve either the entry t in position (s, s k + 1) or the entry y 1 in position (l, l) If we consider the system that arises when equating the characteristic polynomial of B to p(x) = x n + r 1 x n 1 + + r n 1 x + r n, a system of equations arises much like the system in (41) with the following adjustments: 1 for each m with 2 < m l, m s+1, add summand y m 1 y 1 to the coefficient of x n m, 2 if l s k, then for each m with 2 < m l, add summand y 1 ty m 1 to the coefficient of x n m k, 3 if l s + 1, then for each m with 2 < m s k + 1, add summand y 1 ty m 1 to the coefficient of x n m k, and 4 if s k + 1 l s, then remove summand y 1 t from the coefficient of x n k 1 Note that the variable t is not removed from any coefficient of x n m with m a multiple

Sparse spectrally arbitrary patterns 97 of k Thus, the corresponding equation to (42) still provides a monic polynomial in t of odd degree when s is an odd multiple of k Therefore, as in the proof of Theorem 45, the system is solvable by forward substitution if s is an odd multiple of k > 1 In Section 2 we saw that any subdiagonal slide of the pattern C n is spectrally arbitrary and noted that this is not true, in general, for the Y n (s, k) patterns In particular, moving entry (4, 1) of the spectrally arbitrary pattern Y 6 (6, 2) to position (5, 2) produces a pattern that requires singularity We now present a subdiagonal slide that preserves the spectrally arbitrary property of Y n (s, k) Theorem 48 If s is an odd multiple of k > 1, then for s m < n, any m th -subdiagonal slide of Y n (s, k) is spectrally arbitrary Proof Suppose s is an odd multiple of k > 1, and s m < n Suppose B is an m th -subdiagonal slide of Y n (s, k) such that B l,l m+1 = for some l, m < l n Let Y = Y n (s, k) and let B Q(B) be a matrix such B i,j = Y i,j for all i, j except B l,l m+1 = y m and B m,1 = 0 Note the digraph of B has two k-cycles, no s-cycle, and exactly one q-cycle for q {s, k}, 1 q n The composite q-cycles that are not proper q -cycles involve either the entry t in position (s, s k + 1) or the entry y m in position (l, l m + 1) When equating the characteristic polynomial of B to p(x) = x n +r 1 x n 1 + +r n 1 x+r n, a system of equations arises much like the system in (41) except the q th equation may contain some extra terms for all s < q n In order to solve this system we make the same substitutions for the first s equations as we did in the proof for Theorem 45 so that t, y 1,, y s 1 are fixed The remaining n s equations are solvable via forward substitution 5 Concluding comments Of the {0,, }-patterns we observed, including those in this paper and those found via a computer search of small order patterns, the digraph of each irreducible spectrally arbitrary pattern with 2n 1 nonzero entries contains a directed Hamilton path (ie, a directed path on n vertices) In other words, any such pattern is equivalent to a superpattern of a proper Hessenberg pattern (ie, a pattern with a nonzero entry in each position (i, i + 1), 1 i < n) We wonder if there exist examples that do not have a directed Hamilton path or if a Hamilton path is a characteristic of the digraph of every irreducible sparse spectrally arbitrary {0,, }-pattern REFERENCES [1] MS Cavers, SM Fallat Allow problems concerning spectral properties of patterns Electron J Linear Algebra, 23:731 754, 2012 [2] L Corpuz and JJ McDonald Spectrally arbitrary zero-nonzero patterns of order 4 Linear Multilinear Algebra, 55:249 274, 2007 [3] LM DeAlba, IR Hentzel, L Hogben, JJ McDonald, R Mikkelson, O Pryporova, BL Shader

98 B Eastman, B Shader, and KN Vander Meulen and KN Vander Meulen Spectrally arbitrary patterns: Reducibility and the 2n conjecture for n = 5 Linear Algebra Appl, 423:262 276, 2007 [4] JH Drew, CR Johnson, DD Olesky and P van den Driessche Spectrally arbitrary patterns Linear Algebra Appl, 308:121 137, 2000 [5] B Eastman, I-J Kim, BL Shader and KN Vander Meulen Companion matrix patterns Linear Algebra Appl, 463:255 272, 2014 [6] C Eschenbach and Z Li Potentially nilpotent sign pattern matrices Linear Algebra Appl, 299:81 99, 1999 [7] C Garnett and BL Shader The Nilpotent-Centralizer method for spectrally arbitrary patterns Linear Algebra Appl, 438:3836 3850, 2013 [8] JJ McDonald and AA Yielding Complex spectrally arbitrary zero-nonzero patterns Linear Multilinear Algebra, 60:11 26, 2012 [9] AA Yielding Spectrally arbitrary zero-nonzero patterns PhD Thesis, Washington State University, 2009